//
//===----------------------------------------------------------------------===//
-#include "InstCombine.h"
+#include "InstCombineInternal.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/VectorUtils.h"
+#include "llvm/IR/PatternMatch.h"
using namespace llvm;
+using namespace PatternMatch;
-/// CheapToScalarize - Return true if the value is cheaper to scalarize than it
-/// is to leave as a vector operation. isConstant indicates whether we're
-/// extracting one known element. If false we're extracting a variable index.
-static bool CheapToScalarize(Value *V, bool isConstant) {
- if (isa<ConstantAggregateZero>(V))
- return true;
- if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
+#define DEBUG_TYPE "instcombine"
+
+/// Return true if the value is cheaper to scalarize than it is to leave as a
+/// vector operation. isConstant indicates whether we're extracting one known
+/// element. If false we're extracting a variable index.
+static bool cheapToScalarize(Value *V, bool isConstant) {
+ if (Constant *C = dyn_cast<Constant>(V)) {
if (isConstant) return true;
- // If all elts are the same, we can extract.
- Constant *Op0 = C->getOperand(0);
- for (unsigned i = 1; i < C->getNumOperands(); ++i)
- if (C->getOperand(i) != Op0)
- return false;
- return true;
+
+ // If all elts are the same, we can extract it and use any of the values.
+ if (Constant *Op0 = C->getAggregateElement(0U)) {
+ for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e;
+ ++i)
+ if (C->getAggregateElement(i) != Op0)
+ return false;
+ return true;
+ }
}
Instruction *I = dyn_cast<Instruction>(V);
if (!I) return false;
return true;
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
if (BO->hasOneUse() &&
- (CheapToScalarize(BO->getOperand(0), isConstant) ||
- CheapToScalarize(BO->getOperand(1), isConstant)))
+ (cheapToScalarize(BO->getOperand(0), isConstant) ||
+ cheapToScalarize(BO->getOperand(1), isConstant)))
return true;
if (CmpInst *CI = dyn_cast<CmpInst>(I))
if (CI->hasOneUse() &&
- (CheapToScalarize(CI->getOperand(0), isConstant) ||
- CheapToScalarize(CI->getOperand(1), isConstant)))
+ (cheapToScalarize(CI->getOperand(0), isConstant) ||
+ cheapToScalarize(CI->getOperand(1), isConstant)))
return true;
return false;
}
-/// getShuffleMask - Read and decode a shufflevector mask.
-/// Turn undef elements into negative values.
-static SmallVector<int, 16> getShuffleMask(const ShuffleVectorInst *SVI) {
- unsigned NElts = SVI->getType()->getNumElements();
- if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
- return SmallVector<int, 16>(NElts, 0);
- if (isa<UndefValue>(SVI->getOperand(2)))
- return SmallVector<int, 16>(NElts, -1);
-
- SmallVector<int, 16> Result;
- const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
- for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
- if (isa<UndefValue>(*i))
- Result.push_back(-1); // undef
- else
- Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
- return Result;
-}
+// If we have a PHI node with a vector type that has only 2 uses: feed
+// itself and be an operand of extractelement at a constant location,
+// try to replace the PHI of the vector type with a PHI of a scalar type.
+Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
+ // Verify that the PHI node has exactly 2 uses. Otherwise return NULL.
+ if (!PN->hasNUses(2))
+ return nullptr;
+
+ // If so, it's known at this point that one operand is PHI and the other is
+ // an extractelement node. Find the PHI user that is not the extractelement
+ // node.
+ auto iu = PN->user_begin();
+ Instruction *PHIUser = dyn_cast<Instruction>(*iu);
+ if (PHIUser == cast<Instruction>(&EI))
+ PHIUser = cast<Instruction>(*(++iu));
+
+ // Verify that this PHI user has one use, which is the PHI itself,
+ // and that it is a binary operation which is cheap to scalarize.
+ // otherwise return NULL.
+ if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) ||
+ !(isa<BinaryOperator>(PHIUser)) || !cheapToScalarize(PHIUser, true))
+ return nullptr;
+
+ // Create a scalar PHI node that will replace the vector PHI node
+ // just before the current PHI node.
+ PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith(
+ PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN));
+ // Scalarize each PHI operand.
+ for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
+ Value *PHIInVal = PN->getIncomingValue(i);
+ BasicBlock *inBB = PN->getIncomingBlock(i);
+ Value *Elt = EI.getIndexOperand();
+ // If the operand is the PHI induction variable:
+ if (PHIInVal == PHIUser) {
+ // Scalarize the binary operation. Its first operand is the
+ // scalar PHI, and the second operand is extracted from the other
+ // vector operand.
+ BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
+ unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
+ Value *Op = InsertNewInstWith(
+ ExtractElementInst::Create(B0->getOperand(opId), Elt,
+ B0->getOperand(opId)->getName() + ".Elt"),
+ *B0);
+ Value *newPHIUser = InsertNewInstWith(
+ BinaryOperator::Create(B0->getOpcode(), scalarPHI, Op), *B0);
+ scalarPHI->addIncoming(newPHIUser, inBB);
+ } else {
+ // Scalarize PHI input:
+ Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, "");
+ // Insert the new instruction into the predecessor basic block.
+ Instruction *pos = dyn_cast<Instruction>(PHIInVal);
+ BasicBlock::iterator InsertPos;
+ if (pos && !isa<PHINode>(pos)) {
+ InsertPos = ++pos->getIterator();
+ } else {
+ InsertPos = inBB->getFirstInsertionPt();
+ }
-/// FindScalarElement - Given a vector and an element number, see if the scalar
-/// value is already around as a register, for example if it were inserted then
-/// extracted from the vector.
-static Value *FindScalarElement(Value *V, unsigned EltNo) {
- assert(V->getType()->isVectorTy() && "Not looking at a vector?");
- VectorType *PTy = cast<VectorType>(V->getType());
- unsigned Width = PTy->getNumElements();
- if (EltNo >= Width) // Out of range access.
- return UndefValue::get(PTy->getElementType());
-
- if (isa<UndefValue>(V))
- return UndefValue::get(PTy->getElementType());
- if (isa<ConstantAggregateZero>(V))
- return Constant::getNullValue(PTy->getElementType());
- if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
- return CP->getOperand(EltNo);
-
- if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
- // If this is an insert to a variable element, we don't know what it is.
- if (!isa<ConstantInt>(III->getOperand(2)))
- return 0;
- unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
-
- // If this is an insert to the element we are looking for, return the
- // inserted value.
- if (EltNo == IIElt)
- return III->getOperand(1);
-
- // Otherwise, the insertelement doesn't modify the value, recurse on its
- // vector input.
- return FindScalarElement(III->getOperand(0), EltNo);
- }
-
- if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
- unsigned LHSWidth =
- cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
- int InEl = SVI->getMaskValue(EltNo);
- if (InEl < 0)
- return UndefValue::get(PTy->getElementType());
- if (InEl < (int)LHSWidth)
- return FindScalarElement(SVI->getOperand(0), InEl);
- return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
- }
-
- // Otherwise, we don't know.
- return 0;
+ InsertNewInstWith(newEI, *InsertPos);
+
+ scalarPHI->addIncoming(newEI, inBB);
+ }
+ }
+ return ReplaceInstUsesWith(EI, scalarPHI);
}
Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
- // If vector val is undef, replace extract with scalar undef.
- if (isa<UndefValue>(EI.getOperand(0)))
- return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
-
- // If vector val is constant 0, replace extract with scalar 0.
- if (isa<ConstantAggregateZero>(EI.getOperand(0)))
- return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
-
- if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
- // If vector val is constant with all elements the same, replace EI with
- // that element. When the elements are not identical, we cannot replace yet
- // (we do that below, but only when the index is constant).
- Constant *op0 = C->getOperand(0);
- for (unsigned i = 1; i != C->getNumOperands(); ++i)
- if (C->getOperand(i) != op0) {
- op0 = 0;
- break;
- }
- if (op0)
- return ReplaceInstUsesWith(EI, op0);
- }
+ if (Value *V = SimplifyExtractElementInst(
+ EI.getVectorOperand(), EI.getIndexOperand(), DL, TLI, DT, AC))
+ return ReplaceInstUsesWith(EI, V);
+
+ // If vector val is constant with all elements the same, replace EI with
+ // that element. We handle a known element # below.
+ if (Constant *C = dyn_cast<Constant>(EI.getOperand(0)))
+ if (cheapToScalarize(C, false))
+ return ReplaceInstUsesWith(EI, C->getAggregateElement(0U));
// If extracting a specified index from the vector, see if we can recursively
// find a previously computed scalar that was inserted into the vector.
unsigned IndexVal = IdxC->getZExtValue();
unsigned VectorWidth = EI.getVectorOperandType()->getNumElements();
- // If this is extracting an invalid index, turn this into undef, to avoid
- // crashing the code below.
- if (IndexVal >= VectorWidth)
- return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
+ // InstSimplify handles cases where the index is invalid.
+ assert(IndexVal < VectorWidth);
// This instruction only demands the single element from the input vector.
// If the input vector has a single use, simplify it based on this use
APInt UndefElts(VectorWidth, 0);
APInt DemandedMask(VectorWidth, 0);
DemandedMask.setBit(IndexVal);
- if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
- DemandedMask, UndefElts)) {
+ if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), DemandedMask,
+ UndefElts)) {
EI.setOperand(0, V);
return &EI;
}
}
- if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
- return ReplaceInstUsesWith(EI, Elt);
-
- // If the this extractelement is directly using a bitcast from a vector of
+ // If this extractelement is directly using a bitcast from a vector of
// the same number of elements, see if we can find the source element from
// it. In this case, we will end up needing to bitcast the scalars.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
- if (VectorType *VT =
- dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
+ if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
if (VT->getNumElements() == VectorWidth)
- if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
+ if (Value *Elt = findScalarElement(BCI->getOperand(0), IndexVal))
return new BitCastInst(Elt, EI.getType());
}
+
+ // If there's a vector PHI feeding a scalar use through this extractelement
+ // instruction, try to scalarize the PHI.
+ if (PHINode *PN = dyn_cast<PHINode>(EI.getOperand(0))) {
+ Instruction *scalarPHI = scalarizePHI(EI, PN);
+ if (scalarPHI)
+ return scalarPHI;
+ }
}
if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
// Push extractelement into predecessor operation if legal and
- // profitable to do so
+ // profitable to do so.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
if (I->hasOneUse() &&
- CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
+ cheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) {
Value *newEI0 =
Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
EI.getName()+".lhs");
int SrcIdx = SVI->getMaskValue(Elt->getZExtValue());
Value *Src;
unsigned LHSWidth =
- cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
+ SVI->getOperand(0)->getType()->getVectorNumElements();
if (SrcIdx < 0)
return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
SrcIdx, false));
}
} else if (CastInst *CI = dyn_cast<CastInst>(I)) {
- // Canonicalize extractelement(cast) -> cast(extractelement)
- // bitcasts can change the number of vector elements and they cost nothing
- if (CI->hasOneUse() && EI.hasOneUse() &&
- (CI->getOpcode() != Instruction::BitCast)) {
+ // Canonicalize extractelement(cast) -> cast(extractelement).
+ // Bitcasts can change the number of vector elements, and they cost
+ // nothing.
+ if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) {
Value *EE = Builder->CreateExtractElement(CI->getOperand(0),
EI.getIndexOperand());
+ Worklist.AddValue(EE);
return CastInst::Create(CI->getOpcode(), EE, EI.getType());
}
+ } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
+ if (SI->hasOneUse()) {
+ // TODO: For a select on vectors, it might be useful to do this if it
+ // has multiple extractelement uses. For vector select, that seems to
+ // fight the vectorizer.
+
+ // If we are extracting an element from a vector select or a select on
+ // vectors, create a select on the scalars extracted from the vector
+ // arguments.
+ Value *TrueVal = SI->getTrueValue();
+ Value *FalseVal = SI->getFalseValue();
+
+ Value *Cond = SI->getCondition();
+ if (Cond->getType()->isVectorTy()) {
+ Cond = Builder->CreateExtractElement(Cond,
+ EI.getIndexOperand(),
+ Cond->getName() + ".elt");
+ }
+
+ Value *V1Elem
+ = Builder->CreateExtractElement(TrueVal,
+ EI.getIndexOperand(),
+ TrueVal->getName() + ".elt");
+
+ Value *V2Elem
+ = Builder->CreateExtractElement(FalseVal,
+ EI.getIndexOperand(),
+ FalseVal->getName() + ".elt");
+ return SelectInst::Create(Cond,
+ V1Elem,
+ V2Elem,
+ SI->getName() + ".elt");
+ }
}
}
- return 0;
+ return nullptr;
}
-/// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
-/// elements from either LHS or RHS, return the shuffle mask and true.
-/// Otherwise, return false.
-static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
- std::vector<Constant*> &Mask) {
- assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
+/// If V is a shuffle of values that ONLY returns elements from either LHS or
+/// RHS, return the shuffle mask and true. Otherwise, return false.
+static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
+ SmallVectorImpl<Constant*> &Mask) {
+ assert(LHS->getType() == RHS->getType() &&
"Invalid CollectSingleShuffleElements");
- unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
+ unsigned NumElts = V->getType()->getVectorNumElements();
if (isa<UndefValue>(V)) {
Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
- // Okay, we can handle this if the vector we are insertinting into is
+ // We can handle this if the vector we are inserting into is
// transitively ok.
- if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
+ if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
// If so, update the mask to reflect the inserted undef.
Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext()));
return true;
}
} else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
- if (isa<ConstantInt>(EI->getOperand(1)) &&
- EI->getOperand(0)->getType() == V->getType()) {
+ if (isa<ConstantInt>(EI->getOperand(1))) {
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
+ unsigned NumLHSElts = LHS->getType()->getVectorNumElements();
// This must be extracting from either LHS or RHS.
if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
- // Okay, we can handle this if the vector we are insertinting into is
+ // We can handle this if the vector we are inserting into is
// transitively ok.
- if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
+ if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
// If so, update the mask to reflect the inserted value.
if (EI->getOperand(0) == LHS) {
Mask[InsertedIdx % NumElts] =
assert(EI->getOperand(0) == RHS);
Mask[InsertedIdx % NumElts] =
ConstantInt::get(Type::getInt32Ty(V->getContext()),
- ExtractedIdx+NumElts);
+ ExtractedIdx + NumLHSElts);
}
return true;
}
}
}
}
- // TODO: Handle shufflevector here!
return false;
}
-/// CollectShuffleElements - We are building a shuffle of V, using RHS as the
-/// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
-/// that computes V and the LHS value of the shuffle.
-static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
- Value *&RHS) {
- assert(V->getType()->isVectorTy() &&
- (RHS == 0 || V->getType() == RHS->getType()) &&
- "Invalid shuffle!");
+/// If we have insertion into a vector that is wider than the vector that we
+/// are extracting from, try to widen the source vector to allow a single
+/// shufflevector to replace one or more insert/extract pairs.
+static void replaceExtractElements(InsertElementInst *InsElt,
+ ExtractElementInst *ExtElt,
+ InstCombiner &IC) {
+ VectorType *InsVecType = InsElt->getType();
+ VectorType *ExtVecType = ExtElt->getVectorOperandType();
+ unsigned NumInsElts = InsVecType->getVectorNumElements();
+ unsigned NumExtElts = ExtVecType->getVectorNumElements();
+
+ // The inserted-to vector must be wider than the extracted-from vector.
+ if (InsVecType->getElementType() != ExtVecType->getElementType() ||
+ NumExtElts >= NumInsElts)
+ return;
+
+ // Create a shuffle mask to widen the extended-from vector using undefined
+ // values. The mask selects all of the values of the original vector followed
+ // by as many undefined values as needed to create a vector of the same length
+ // as the inserted-to vector.
+ SmallVector<Constant *, 16> ExtendMask;
+ IntegerType *IntType = Type::getInt32Ty(InsElt->getContext());
+ for (unsigned i = 0; i < NumExtElts; ++i)
+ ExtendMask.push_back(ConstantInt::get(IntType, i));
+ for (unsigned i = NumExtElts; i < NumInsElts; ++i)
+ ExtendMask.push_back(UndefValue::get(IntType));
+
+ Value *ExtVecOp = ExtElt->getVectorOperand();
+ auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
+ ConstantVector::get(ExtendMask));
+
+ // Insert the new shuffle after the vector operand of the extract is defined
+ // (as long as it's not a PHI) or at the start of the basic block of the
+ // extract, so any subsequent extracts in the same basic block can use it.
+ // TODO: Insert before the earliest ExtractElementInst that is replaced.
+ auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp);
+ if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst))
+ WideVec->insertAfter(ExtVecOpInst);
+ else
+ IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
+
+ // Replace extracts from the original narrow vector with extracts from the new
+ // wide vector.
+ for (User *U : ExtVecOp->users()) {
+ ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
+ if (!OldExt || OldExt->getParent() != WideVec->getParent())
+ continue;
+ auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
+ NewExt->insertAfter(WideVec);
+ IC.ReplaceInstUsesWith(*OldExt, NewExt);
+ }
+}
+
+/// We are building a shuffle to create V, which is a sequence of insertelement,
+/// extractelement pairs. If PermittedRHS is set, then we must either use it or
+/// not rely on the second vector source. Return a std::pair containing the
+/// left and right vectors of the proposed shuffle (or 0), and set the Mask
+/// parameter as required.
+///
+/// Note: we intentionally don't try to fold earlier shuffles since they have
+/// often been chosen carefully to be efficiently implementable on the target.
+typedef std::pair<Value *, Value *> ShuffleOps;
+
+static ShuffleOps collectShuffleElements(Value *V,
+ SmallVectorImpl<Constant *> &Mask,
+ Value *PermittedRHS,
+ InstCombiner &IC) {
+ assert(V->getType()->isVectorTy() && "Invalid shuffle!");
unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
if (isa<UndefValue>(V)) {
Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext())));
- return V;
+ return std::make_pair(
+ PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr);
}
-
+
if (isa<ConstantAggregateZero>(V)) {
Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0));
- return V;
+ return std::make_pair(V, nullptr);
}
-
+
if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
// If this is an insert of an extract from some other vector, include it.
Value *VecOp = IEI->getOperand(0);
Value *IdxOp = IEI->getOperand(2);
if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
- if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
- EI->getOperand(0)->getType() == V->getType()) {
+ if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
// Either the extracted from or inserted into vector must be RHSVec,
// otherwise we'd end up with a shuffle of three inputs.
- if (EI->getOperand(0) == RHS || RHS == 0) {
- RHS = EI->getOperand(0);
- Value *V = CollectShuffleElements(VecOp, Mask, RHS);
+ if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) {
+ Value *RHS = EI->getOperand(0);
+ ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC);
+ assert(LR.second == nullptr || LR.second == RHS);
+
+ if (LR.first->getType() != RHS->getType()) {
+ // Although we are giving up for now, see if we can create extracts
+ // that match the inserts for another round of combining.
+ replaceExtractElements(IEI, EI, IC);
+
+ // We tried our best, but we can't find anything compatible with RHS
+ // further up the chain. Return a trivial shuffle.
+ for (unsigned i = 0; i < NumElts; ++i)
+ Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), i);
+ return std::make_pair(V, nullptr);
+ }
+
+ unsigned NumLHSElts = RHS->getType()->getVectorNumElements();
Mask[InsertedIdx % NumElts] =
ConstantInt::get(Type::getInt32Ty(V->getContext()),
- NumElts+ExtractedIdx);
- return V;
+ NumLHSElts+ExtractedIdx);
+ return std::make_pair(LR.first, RHS);
}
- if (VecOp == RHS) {
- Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
- // Everything but the extracted element is replaced with the RHS.
- for (unsigned i = 0; i != NumElts; ++i) {
- if (i != InsertedIdx)
- Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()),
- NumElts+i);
- }
- return V;
+ if (VecOp == PermittedRHS) {
+ // We've gone as far as we can: anything on the other side of the
+ // extractelement will already have been converted into a shuffle.
+ unsigned NumLHSElts =
+ EI->getOperand(0)->getType()->getVectorNumElements();
+ for (unsigned i = 0; i != NumElts; ++i)
+ Mask.push_back(ConstantInt::get(
+ Type::getInt32Ty(V->getContext()),
+ i == InsertedIdx ? ExtractedIdx : NumLHSElts + i));
+ return std::make_pair(EI->getOperand(0), PermittedRHS);
}
// If this insertelement is a chain that comes from exactly these two
// vectors, return the vector and the effective shuffle.
- if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
- return EI->getOperand(0);
+ if (EI->getOperand(0)->getType() == PermittedRHS->getType() &&
+ collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS,
+ Mask))
+ return std::make_pair(EI->getOperand(0), PermittedRHS);
}
}
}
- // TODO: Handle shufflevector here!
- // Otherwise, can't do anything fancy. Return an identity vector.
+ // Otherwise, we can't do anything fancy. Return an identity vector.
for (unsigned i = 0; i != NumElts; ++i)
Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i));
- return V;
+ return std::make_pair(V, nullptr);
+}
+
+/// Try to find redundant insertvalue instructions, like the following ones:
+/// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
+/// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
+/// Here the second instruction inserts values at the same indices, as the
+/// first one, making the first one redundant.
+/// It should be transformed to:
+/// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
+Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
+ bool IsRedundant = false;
+ ArrayRef<unsigned int> FirstIndices = I.getIndices();
+
+ // If there is a chain of insertvalue instructions (each of them except the
+ // last one has only one use and it's another insertvalue insn from this
+ // chain), check if any of the 'children' uses the same indices as the first
+ // instruction. In this case, the first one is redundant.
+ Value *V = &I;
+ unsigned Depth = 0;
+ while (V->hasOneUse() && Depth < 10) {
+ User *U = V->user_back();
+ auto UserInsInst = dyn_cast<InsertValueInst>(U);
+ if (!UserInsInst || U->getOperand(0) != V)
+ break;
+ if (UserInsInst->getIndices() == FirstIndices) {
+ IsRedundant = true;
+ break;
+ }
+ V = UserInsInst;
+ Depth++;
+ }
+
+ if (IsRedundant)
+ return ReplaceInstUsesWith(I, I.getOperand(0));
+ return nullptr;
}
Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
// If the inserted element was extracted from some other vector, and if the
// indexes are constant, try to turn this into a shufflevector operation.
if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
- if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
- EI->getOperand(0)->getType() == IE.getType()) {
- unsigned NumVectorElts = IE.getType()->getNumElements();
+ if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) {
+ unsigned NumInsertVectorElts = IE.getType()->getNumElements();
+ unsigned NumExtractVectorElts =
+ EI->getOperand(0)->getType()->getVectorNumElements();
unsigned ExtractedIdx =
cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
- if (ExtractedIdx >= NumVectorElts) // Out of range extract.
+ if (ExtractedIdx >= NumExtractVectorElts) // Out of range extract.
return ReplaceInstUsesWith(IE, VecOp);
- if (InsertedIdx >= NumVectorElts) // Out of range insert.
+ if (InsertedIdx >= NumInsertVectorElts) // Out of range insert.
return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
// If we are extracting a value from a vector, then inserting it right
// If this insertelement isn't used by some other insertelement, turn it
// (and any insertelements it points to), into one big shuffle.
- if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
- std::vector<Constant*> Mask;
- Value *RHS = 0;
- Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
- if (RHS == 0) RHS = UndefValue::get(LHS->getType());
- // We now have a shuffle of LHS, RHS, Mask.
- return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
+ if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.user_back())) {
+ SmallVector<Constant*, 16> Mask;
+ ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this);
+
+ // The proposed shuffle may be trivial, in which case we shouldn't
+ // perform the combine.
+ if (LR.first != &IE && LR.second != &IE) {
+ // We now have a shuffle of LHS, RHS, Mask.
+ if (LR.second == nullptr)
+ LR.second = UndefValue::get(LR.first->getType());
+ return new ShuffleVectorInst(LR.first, LR.second,
+ ConstantVector::get(Mask));
+ }
}
}
}
return &IE;
}
- return 0;
+ return nullptr;
+}
+
+/// Return true if we can evaluate the specified expression tree if the vector
+/// elements were shuffled in a different order.
+static bool CanEvaluateShuffled(Value *V, ArrayRef<int> Mask,
+ unsigned Depth = 5) {
+ // We can always reorder the elements of a constant.
+ if (isa<Constant>(V))
+ return true;
+
+ // We won't reorder vector arguments. No IPO here.
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false;
+
+ // Two users may expect different orders of the elements. Don't try it.
+ if (!I->hasOneUse())
+ return false;
+
+ if (Depth == 0) return false;
+
+ switch (I->getOpcode()) {
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::UIToFP:
+ case Instruction::SIToFP:
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::GetElementPtr: {
+ for (Value *Operand : I->operands()) {
+ if (!CanEvaluateShuffled(Operand, Mask, Depth-1))
+ return false;
+ }
+ return true;
+ }
+ case Instruction::InsertElement: {
+ ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2));
+ if (!CI) return false;
+ int ElementNumber = CI->getLimitedValue();
+
+ // Verify that 'CI' does not occur twice in Mask. A single 'insertelement'
+ // can't put an element into multiple indices.
+ bool SeenOnce = false;
+ for (int i = 0, e = Mask.size(); i != e; ++i) {
+ if (Mask[i] == ElementNumber) {
+ if (SeenOnce)
+ return false;
+ SeenOnce = true;
+ }
+ }
+ return CanEvaluateShuffled(I->getOperand(0), Mask, Depth-1);
+ }
+ }
+ return false;
+}
+
+/// Rebuild a new instruction just like 'I' but with the new operands given.
+/// In the event of type mismatch, the type of the operands is correct.
+static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) {
+ // We don't want to use the IRBuilder here because we want the replacement
+ // instructions to appear next to 'I', not the builder's insertion point.
+ switch (I->getOpcode()) {
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor: {
+ BinaryOperator *BO = cast<BinaryOperator>(I);
+ assert(NewOps.size() == 2 && "binary operator with #ops != 2");
+ BinaryOperator *New =
+ BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
+ NewOps[0], NewOps[1], "", BO);
+ if (isa<OverflowingBinaryOperator>(BO)) {
+ New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
+ New->setHasNoSignedWrap(BO->hasNoSignedWrap());
+ }
+ if (isa<PossiblyExactOperator>(BO)) {
+ New->setIsExact(BO->isExact());
+ }
+ if (isa<FPMathOperator>(BO))
+ New->copyFastMathFlags(I);
+ return New;
+ }
+ case Instruction::ICmp:
+ assert(NewOps.size() == 2 && "icmp with #ops != 2");
+ return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
+ NewOps[0], NewOps[1]);
+ case Instruction::FCmp:
+ assert(NewOps.size() == 2 && "fcmp with #ops != 2");
+ return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
+ NewOps[0], NewOps[1]);
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::UIToFP:
+ case Instruction::SIToFP:
+ case Instruction::FPTrunc:
+ case Instruction::FPExt: {
+ // It's possible that the mask has a different number of elements from
+ // the original cast. We recompute the destination type to match the mask.
+ Type *DestTy =
+ VectorType::get(I->getType()->getScalarType(),
+ NewOps[0]->getType()->getVectorNumElements());
+ assert(NewOps.size() == 1 && "cast with #ops != 1");
+ return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
+ "", I);
+ }
+ case Instruction::GetElementPtr: {
+ Value *Ptr = NewOps[0];
+ ArrayRef<Value*> Idx = NewOps.slice(1);
+ GetElementPtrInst *GEP = GetElementPtrInst::Create(
+ cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I);
+ GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
+ return GEP;
+ }
+ }
+ llvm_unreachable("failed to rebuild vector instructions");
+}
+
+Value *
+InstCombiner::EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) {
+ // Mask.size() does not need to be equal to the number of vector elements.
+
+ assert(V->getType()->isVectorTy() && "can't reorder non-vector elements");
+ if (isa<UndefValue>(V)) {
+ return UndefValue::get(VectorType::get(V->getType()->getScalarType(),
+ Mask.size()));
+ }
+ if (isa<ConstantAggregateZero>(V)) {
+ return ConstantAggregateZero::get(
+ VectorType::get(V->getType()->getScalarType(),
+ Mask.size()));
+ }
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ SmallVector<Constant *, 16> MaskValues;
+ for (int i = 0, e = Mask.size(); i != e; ++i) {
+ if (Mask[i] == -1)
+ MaskValues.push_back(UndefValue::get(Builder->getInt32Ty()));
+ else
+ MaskValues.push_back(Builder->getInt32(Mask[i]));
+ }
+ return ConstantExpr::getShuffleVector(C, UndefValue::get(C->getType()),
+ ConstantVector::get(MaskValues));
+ }
+
+ Instruction *I = cast<Instruction>(V);
+ switch (I->getOpcode()) {
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::UIToFP:
+ case Instruction::SIToFP:
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::Select:
+ case Instruction::GetElementPtr: {
+ SmallVector<Value*, 8> NewOps;
+ bool NeedsRebuild = (Mask.size() != I->getType()->getVectorNumElements());
+ for (int i = 0, e = I->getNumOperands(); i != e; ++i) {
+ Value *V = EvaluateInDifferentElementOrder(I->getOperand(i), Mask);
+ NewOps.push_back(V);
+ NeedsRebuild |= (V != I->getOperand(i));
+ }
+ if (NeedsRebuild) {
+ return buildNew(I, NewOps);
+ }
+ return I;
+ }
+ case Instruction::InsertElement: {
+ int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue();
+
+ // The insertelement was inserting at Element. Figure out which element
+ // that becomes after shuffling. The answer is guaranteed to be unique
+ // by CanEvaluateShuffled.
+ bool Found = false;
+ int Index = 0;
+ for (int e = Mask.size(); Index != e; ++Index) {
+ if (Mask[Index] == Element) {
+ Found = true;
+ break;
+ }
+ }
+
+ // If element is not in Mask, no need to handle the operand 1 (element to
+ // be inserted). Just evaluate values in operand 0 according to Mask.
+ if (!Found)
+ return EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
+
+ Value *V = EvaluateInDifferentElementOrder(I->getOperand(0), Mask);
+ return InsertElementInst::Create(V, I->getOperand(1),
+ Builder->getInt32(Index), "", I);
+ }
+ }
+ llvm_unreachable("failed to reorder elements of vector instruction!");
}
+static void recognizeIdentityMask(const SmallVectorImpl<int> &Mask,
+ bool &isLHSID, bool &isRHSID) {
+ isLHSID = isRHSID = true;
+
+ for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
+ if (Mask[i] < 0) continue; // Ignore undef values.
+ // Is this an identity shuffle of the LHS value?
+ isLHSID &= (Mask[i] == (int)i);
+
+ // Is this an identity shuffle of the RHS value?
+ isRHSID &= (Mask[i]-e == i);
+ }
+}
+
+// Returns true if the shuffle is extracting a contiguous range of values from
+// LHS, for example:
+// +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+// Input: |AA|BB|CC|DD|EE|FF|GG|HH|II|JJ|KK|LL|MM|NN|OO|PP|
+// Shuffles to: |EE|FF|GG|HH|
+// +--+--+--+--+
+static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI,
+ SmallVector<int, 16> &Mask) {
+ unsigned LHSElems =
+ cast<VectorType>(SVI.getOperand(0)->getType())->getNumElements();
+ unsigned MaskElems = Mask.size();
+ unsigned BegIdx = Mask.front();
+ unsigned EndIdx = Mask.back();
+ if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1)
+ return false;
+ for (unsigned I = 0; I != MaskElems; ++I)
+ if (static_cast<unsigned>(Mask[I]) != BegIdx + I)
+ return false;
+ return true;
+}
Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
Value *LHS = SVI.getOperand(0);
Value *RHS = SVI.getOperand(1);
- SmallVector<int, 16> Mask = getShuffleMask(&SVI);
+ SmallVector<int, 16> Mask = SVI.getShuffleMask();
+ Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
bool MadeChange = false;
if (LHS == RHS || isa<UndefValue>(LHS)) {
if (isa<UndefValue>(LHS) && LHS == RHS) {
// shuffle(undef,undef,mask) -> undef.
- Value* result = (VWidth == LHSWidth)
+ Value *Result = (VWidth == LHSWidth)
? LHS : UndefValue::get(SVI.getType());
- return ReplaceInstUsesWith(SVI, result);
+ return ReplaceInstUsesWith(SVI, Result);
}
// Remap any references to RHS to use LHS.
- std::vector<Constant*> Elts;
+ SmallVector<Constant*, 16> Elts;
for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) {
- if (Mask[i] < 0)
- Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
- else {
- if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
- (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
- Mask[i] = -1; // Turn into undef.
- Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext())));
- } else {
- Mask[i] = Mask[i] % e; // Force to LHS.
- Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()),
- Mask[i]));
- }
+ if (Mask[i] < 0) {
+ Elts.push_back(UndefValue::get(Int32Ty));
+ continue;
+ }
+
+ if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) ||
+ (Mask[i] < (int)e && isa<UndefValue>(LHS))) {
+ Mask[i] = -1; // Turn into undef.
+ Elts.push_back(UndefValue::get(Int32Ty));
+ } else {
+ Mask[i] = Mask[i] % e; // Force to LHS.
+ Elts.push_back(ConstantInt::get(Int32Ty, Mask[i]));
}
}
SVI.setOperand(0, SVI.getOperand(1));
if (VWidth == LHSWidth) {
// Analyze the shuffle, are the LHS or RHS and identity shuffles?
- bool isLHSID = true, isRHSID = true;
-
- for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
- if (Mask[i] < 0) continue; // Ignore undef values.
- // Is this an identity shuffle of the LHS value?
- isLHSID &= (Mask[i] == (int)i);
-
- // Is this an identity shuffle of the RHS value?
- isRHSID &= (Mask[i]-e == i);
- }
+ bool isLHSID, isRHSID;
+ recognizeIdentityMask(Mask, isLHSID, isRHSID);
// Eliminate identity shuffles.
if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
}
+ if (isa<UndefValue>(RHS) && CanEvaluateShuffled(LHS, Mask)) {
+ Value *V = EvaluateInDifferentElementOrder(LHS, Mask);
+ return ReplaceInstUsesWith(SVI, V);
+ }
+
+ // SROA generates shuffle+bitcast when the extracted sub-vector is bitcast to
+ // a non-vector type. We can instead bitcast the original vector followed by
+ // an extract of the desired element:
+ //
+ // %sroa = shufflevector <16 x i8> %in, <16 x i8> undef,
+ // <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ // %1 = bitcast <4 x i8> %sroa to i32
+ // Becomes:
+ // %bc = bitcast <16 x i8> %in to <4 x i32>
+ // %ext = extractelement <4 x i32> %bc, i32 0
+ //
+ // If the shuffle is extracting a contiguous range of values from the input
+ // vector then each use which is a bitcast of the extracted size can be
+ // replaced. This will work if the vector types are compatible, and the begin
+ // index is aligned to a value in the casted vector type. If the begin index
+ // isn't aligned then we can shuffle the original vector (keeping the same
+ // vector type) before extracting.
+ //
+ // This code will bail out if the target type is fundamentally incompatible
+ // with vectors of the source type.
+ //
+ // Example of <16 x i8>, target type i32:
+ // Index range [4,8): v-----------v Will work.
+ // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+ // <16 x i8>: | | | | | | | | | | | | | | | | |
+ // <4 x i32>: | | | | |
+ // +-----------+-----------+-----------+-----------+
+ // Index range [6,10): ^-----------^ Needs an extra shuffle.
+ // Target type i40: ^--------------^ Won't work, bail.
+ if (isShuffleExtractingFromLHS(SVI, Mask)) {
+ Value *V = LHS;
+ unsigned MaskElems = Mask.size();
+ unsigned BegIdx = Mask.front();
+ VectorType *SrcTy = cast<VectorType>(V->getType());
+ unsigned VecBitWidth = SrcTy->getBitWidth();
+ unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType());
+ assert(SrcElemBitWidth && "vector elements must have a bitwidth");
+ unsigned SrcNumElems = SrcTy->getNumElements();
+ SmallVector<BitCastInst *, 8> BCs;
+ DenseMap<Type *, Value *> NewBCs;
+ for (User *U : SVI.users())
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(U))
+ if (!BC->use_empty())
+ // Only visit bitcasts that weren't previously handled.
+ BCs.push_back(BC);
+ for (BitCastInst *BC : BCs) {
+ Type *TgtTy = BC->getDestTy();
+ unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy);
+ if (!TgtElemBitWidth)
+ continue;
+ unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth;
+ bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth;
+ bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth);
+ if (!VecBitWidthsEqual)
+ continue;
+ if (!VectorType::isValidElementType(TgtTy))
+ continue;
+ VectorType *CastSrcTy = VectorType::get(TgtTy, TgtNumElems);
+ if (!BegIsAligned) {
+ // Shuffle the input so [0,NumElements) contains the output, and
+ // [NumElems,SrcNumElems) is undef.
+ SmallVector<Constant *, 16> ShuffleMask(SrcNumElems,
+ UndefValue::get(Int32Ty));
+ for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I)
+ ShuffleMask[I] = ConstantInt::get(Int32Ty, Idx);
+ V = Builder->CreateShuffleVector(V, UndefValue::get(V->getType()),
+ ConstantVector::get(ShuffleMask),
+ SVI.getName() + ".extract");
+ BegIdx = 0;
+ }
+ unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth;
+ assert(SrcElemsPerTgtElem);
+ BegIdx /= SrcElemsPerTgtElem;
+ bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end();
+ auto *NewBC =
+ BCAlreadyExists
+ ? NewBCs[CastSrcTy]
+ : Builder->CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc");
+ if (!BCAlreadyExists)
+ NewBCs[CastSrcTy] = NewBC;
+ auto *Ext = Builder->CreateExtractElement(
+ NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract");
+ // The shufflevector isn't being replaced: the bitcast that used it
+ // is. InstCombine will visit the newly-created instructions.
+ ReplaceInstUsesWith(*BC, Ext);
+ MadeChange = true;
+ }
+ }
+
// If the LHS is a shufflevector itself, see if we can combine it with this
// one without producing an unusual shuffle.
// Cases that might be simplified:
ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
if (LHSShuffle)
if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS))
- LHSShuffle = NULL;
+ LHSShuffle = nullptr;
if (RHSShuffle)
if (!isa<UndefValue>(RHSShuffle->getOperand(1)))
- RHSShuffle = NULL;
+ RHSShuffle = nullptr;
if (!LHSShuffle && !RHSShuffle)
- return MadeChange ? &SVI : 0;
+ return MadeChange ? &SVI : nullptr;
- Value* LHSOp0 = NULL;
- Value* LHSOp1 = NULL;
- Value* RHSOp0 = NULL;
+ Value* LHSOp0 = nullptr;
+ Value* LHSOp1 = nullptr;
+ Value* RHSOp0 = nullptr;
unsigned LHSOp0Width = 0;
unsigned RHSOp0Width = 0;
if (LHSShuffle) {
// case 4
if (LHSOp0 == RHSOp0) {
newLHS = LHSOp0;
- newRHS = NULL;
+ newRHS = nullptr;
}
if (newLHS == LHS && newRHS == RHS)
- return MadeChange ? &SVI : 0;
+ return MadeChange ? &SVI : nullptr;
SmallVector<int, 16> LHSMask;
SmallVector<int, 16> RHSMask;
- if (newLHS != LHS) {
- LHSMask = getShuffleMask(LHSShuffle);
- }
- if (RHSShuffle && newRHS != RHS) {
- RHSMask = getShuffleMask(RHSShuffle);
- }
+ if (newLHS != LHS)
+ LHSMask = LHSShuffle->getShuffleMask();
+ if (RHSShuffle && newRHS != RHS)
+ RHSMask = RHSShuffle->getShuffleMask();
+
unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth;
SmallVector<int, 16> newMask;
bool isSplat = true;
// ShuffleVectorInst is equivalent to the original one.
for (unsigned i = 0; i < VWidth; ++i) {
int eltMask;
- if (Mask[i] == -1) {
+ if (Mask[i] < 0) {
// This element is an undef value.
eltMask = -1;
} else if (Mask[i] < (int)LHSWidth) {
// This element is from left hand side vector operand.
- //
+ //
// If LHS is going to be replaced (case 1, 2, or 4), calculate the
// new mask value for the element.
if (newLHS != LHS) {
// with a -1 mask value.
if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1))
eltMask = -1;
- }
- else
+ } else
eltMask = Mask[i];
} else {
// This element is from right hand side vector operand
&& "should have been check above");
eltMask = -1;
}
- }
- else
+ } else
eltMask = Mask[i]-LHSWidth;
// If LHS's width is changed, shift the mask value accordingly.
// If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any
- // references to RHSOp0 to LHSOp0, so we don't need to shift the mask.
- if (eltMask >= 0 && newRHS != NULL)
+ // references from RHSOp0 to LHSOp0, so we don't need to shift the mask.
+ // If newRHS == newLHS, we want to remap any references from newRHS to
+ // newLHS so that we can properly identify splats that may occur due to
+ // obfuscation across the two vectors.
+ if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS)
eltMask += newLHSWidth;
}
// or is a splat, do the replacement.
if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
SmallVector<Constant*, 16> Elts;
- Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
for (unsigned i = 0, e = newMask.size(); i != e; ++i) {
if (newMask[i] < 0) {
Elts.push_back(UndefValue::get(Int32Ty));
Elts.push_back(ConstantInt::get(Int32Ty, newMask[i]));
}
}
- if (newRHS == NULL)
+ if (!newRHS)
newRHS = UndefValue::get(newLHS->getType());
return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts));
}
- return MadeChange ? &SVI : 0;
+ // If the result mask is an identity, replace uses of this instruction with
+ // corresponding argument.
+ bool isLHSID, isRHSID;
+ recognizeIdentityMask(newMask, isLHSID, isRHSID);
+ if (isLHSID && VWidth == LHSOp0Width) return ReplaceInstUsesWith(SVI, newLHS);
+ if (isRHSID && VWidth == RHSOp0Width) return ReplaceInstUsesWith(SVI, newRHS);
+
+ return MadeChange ? &SVI : nullptr;
}
projects / oota-llvm.git / blobdiff